{"id":"d551c89d-9bac-46ef-ad7d-488debf95ea5","arxiv_id":"2506.21843","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":6,"one_line_summary":"3D-Telepathy reconstructs 3D objects from EEG signals by combining a dual self-attention EEG encoder with stable diffusion and variational score distillation into a NeRF, and reports best 2D-frame metrics among compared EEG decoders.","lead":"This paper trains an EEG encoder to translate brain signals recorded while viewing images into 3D objects, using a diffusion model and neural radiance fields. If it works reliably, it would let people with limited communication generate or select 3D content directly from brain activity.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Evaluation lacks a held-out split and an unconditional control, so the reported 2D-frame PSNR/LPIPS/FID cannot establish that EEG signals, rather than memorization or a strong image prior, drive the 3D reconstructions.","rationale":"The reader's weakest assumption correctly identifies the absence of a held-out split as the main threat to the central claim. My independent read of the full text confirms this: Section IV describes training and evaluation on EEG-image pairs but never reports how the data were partitioned, whether any EEG trials were reserved, or whether subjects were held out. The quantitative evidence in Tables I and II is therefore consistent with in-sample fitting. The paper's ablations vary training components and hyperparameters but do not include an unconditional or shuffled-condition control, so they cannot establish that the EEG embedding itself is what drives the 3D output. If the EEG condition were removed and the metrics stayed similar, the claimed decoding would not be supported. The dataset attribution to MOABB is also a concrete error, but it is secondary to the evaluation gap because it affects reproducibility rather than the logic of the decoding claim. A strict held-out split with an unconditional control is a single, feasible experiment that would decisively test whether the reported PSNR/LPIPS/FID reflect EEG-driven 3D reconstruction or memorization plus a strong generative prior. Until such an experiment is reported, the conditional verdict is appropriate: the paper's central claim is plausible and meaningful but not yet supported by the evidence presented.","tokens_in":11009,"tokens_out":5075,"duration_ms":59247,"concrete_test":"Re-run the complete pipeline (EEG encoder training, U-Net/LoRA fine-tuning, VSD-NeRF generation) using a strict 80/20 held-out split by EEG trial, with no overlap between training and evaluation pairs, and additionally by subject (e.g., leave-one-subject-out). Report PSNR, LPIPS, FID, and top-1 identity/category retrieval on held-out EEGs, and include an unconditional control where the EEG embedding is replaced by a fixed mean embedding or random noise. If held-out metrics and retrieval are at or near the unconditional control, the EEG-conditioning claim fails; if they remain substantially better, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the EEG encoder extracts enough visual-semantic information to condition a NeRF through variational score distillation. For that claim to hold, the model must generalize to EEG signals not seen during training. Section IV.A describes training on EEG-ImageNet pairs, Section IV.B fine-tunes the U-Net on the same EEG-image pairs, and Section IV.D reports PSNR 28.5203, LPIPS 0.3178, and FID 57.3326 after extracting and manually aligning NeRF frames with reference images. No train/test split, subject split, or held-out evaluation is reported anywhere in Section IV, and Tables I and III also appear to use in-sample reconstructions. If the evaluated frames come from EEG trials already used in Stages 1 and 2, the metrics can be inflated by memorization and do not support the generalization claim. A second, related weakness is the absence of any no-condition or shuffled-EEG baseline: without comparing against a random or mean EEG embedding, one cannot tell whether the NeRF output is conditioned on EEG content or on the Stable Diffusion/NeRF prior alone, which would produce plausible images even with no decoding. The dataset citation is also unreliable: reference [12] is the MOABB BCI benchmarking toolbox, not an EEG-ImageNet dataset. The decisive issue, however, is the missing evidence for out-of-sample decoding.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes 3D-Telepathy, an end-to-end pipeline that decodes EEG signals into 3D object representations. It introduces a dual self-attention EEG encoder trained with masked autoencoding, contrastive learning, and cross-attention losses, then uses the resulting embeddings as conditioning for a NeRF optimized through variational score distillation (VSD) with a Stable Diffusion prior. Quantitative evaluation compares rendered NeRF frames against static reference images using PSNR, LPIPS, and FID, and the paper reports ablations over training losses and hyperparameters.","tokens_in":11273,"tokens_out":6121,"duration_ms":56843,"significance":"If validated, this would be a first demonstration of 3D visual reconstruction from EEG signals and would extend cross-modal brain decoding into the 3D domain. Strengths include the use of a standard VSD update rule, the combination of self-supervised and contrastive objectives for the EEG encoder, and the candid acknowledgment of limitations such as loss of fine texture and poor fidelity on unseen object sides. The code link is provided, though its contents were not verified during review. However, the current experiments do not establish out-of-sample decoding: no held-out split, subject-independent evaluation, or control condition is reported, and the dataset provenance is questionable. The central claim is therefore not yet adequately supported.","major_comments":[{"comment":"The evaluation never reports a held-out train/test split, a subject-independent split, or any repeated-run statistics. Section IV.A describes the training data, Section IV.B fine-tunes the U-Net on EEG-image pairs, and Section IV.D reports PSNR/LPIPS/FID on frames aligned with EEG-ImageNet reference images. Because the text does not state that the evaluated stimuli were excluded from Stage 1, Stage 2, and the U-Net fine-tuning, the headline numbers (PSNR 28.5203, LPIPS 0.3178, FID 57.3326) could reflect in-sample memorization rather than generalizable decoding of EEG signals.","section":"Section IV.A, IV.B, IV.D"},{"comment":"There is no control condition: the paper does not compare against unconditional generation, a mean-EEG embedding, or shuffled EEG labels. Since the pipeline conditions a strong Stable Diffusion/NeRF prior with CFG=3, such a baseline is necessary to establish that the EEG embedding, rather than the image prior alone, drives the reconstruction content.","section":"Section III.C and Section IV.D"},{"comment":"The dataset citation is not reliable. Reference [12] is the MOABB BCI benchmarking toolbox by Jayaram and Barachant, not an EEG-ImageNet dataset containing approximately 40 categories, 2000 images, and 128-channel EEG recordings during image viewing. The authors must cite the actual dataset and provide details on subjects, number of trials per image, and the exact train/test protocol.","section":"Section IV.A"},{"comment":"The central 3D reconstruction claim is quantified only by 2D frame metrics after a viewpoint-alignment procedure, and no 3D ground-truth metric (e.g., Chamfer distance, depth accuracy, or volume IoU) is reported. The viewpoint-matching step is not described in a way that rules out manual selection of favorable frames, and the paper itself acknowledges that unseen object sides have poor fidelity; this limitation requires quantitative evaluation.","section":"Section IV.D"},{"comment":"The comparison in Table II is not controlled: CapsEEGNet, Brain2Image, DreamDiffusion, and Dongyang Li et al. are 2D image-reconstruction methods, whereas the proposed method selects 3D-rendered frames for comparison. Differences in PSNR/LPIPS/FID may reflect the evaluation protocol rather than the EEG-decoding capability; results should be compared under the same image-generation setup or at least with the same train/test split.","section":"Section IV.D, Table II"}],"minor_comments":[{"comment":"The abbreviation VSD is expanded as 'Variational Score Diffusion' but the method is Variational Score Distillation; please correct the terminology throughout.","section":"Section III.C"},{"comment":"The checkmarks in Table I do not align unambiguously with the four training components, and the percentage changes quoted in the text (23.7%, 123.9%, 228.7%) do not exactly match the table values; please make the table readable and recompute the percentages.","section":"Section IV.E, Table I"},{"comment":"The phrase 'padded to maintain 128 channels' is unclear when the acquisition device already used 128 electrodes; specify which channels were missing and how the padding was performed.","section":"Section IV.A"},{"comment":"References [23] and [24] are introduced as existing EEG-to-3D works but are not included in the quantitative comparison; please state whether their code or results were unavailable and discuss how this affects the claim of being the first end-to-end EEG-to-3D model.","section":"Section IV.D and References [23], [24]"}],"recommendation":"major_revision","confidential_remarks":"The main risk is the lack of a held-out split and the suspicious dataset citation. I would ask for out-of-sample evaluation on a defined test split, an unconditional or shuffled-EEG baseline, and a corrected dataset reference before considering the paper. The VSD pipeline is standard, so the novelty rests on the EEG encoder and on the evaluation; both need to be strengthened substantially."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know this paper before the hype: it builds a reasonable end-to-end pipeline from EEG to 3D, but the experiments as reported do not establish that the EEG signal, rather than the diffusion prior or in-sample memorization, drives the reconstructions. The central claim is conditional at best.\n\nWhat's new and worth credit: the task itself is real and underexplored. Prior work stopped at EEG-to-2D; extending to NeRF via variational score distillation is a sensible next step. The dual self-attention EEG encoder and the hybrid training recipe (MAE-style masking, contrastive learning, cross-attention with images) are all known ingredients, but the combination is novel and the ablation study in Table I suggests each piece contributes. They also ship a code link and are candid in the conclusion about texture loss and inability to hallucinate unseen views. That honesty earns goodwill.\n\nThe soft spots are serious, though. First, there is no train/test split, no subject-independent evaluation, and no error bars anywhere in Section IV. All reported PSNR/LPIPS/FID numbers appear to be on in-sample reconstructions. If the evaluated EEG trials were used in Stages 1 and 2, those numbers can be inflated by memorization and say nothing about generalization. Second, there is no no-condition or shuffled-EEG baseline. A random or mean embedding fed through the same Stable Diffusion + NeRF prior would likely produce plausible images, so without that control you cannot attribute the output to neural decoding at all. Third, reference [12] is misattributed: MOABB is a BCI benchmarking toolbox, not an EEG-ImageNet dataset. The actual data source, probably an earlier EEG visual decoding dataset, is not properly cited. That is a reproducibility problem. Fourth, judging 3D reconstruction quality by 2D frame metrics after manual viewpoint alignment is clever but indirect; a Chamfer distance or depth comparison against a known 3D mesh would be far more convincing. They do acknowledge the limitation of unseen sides, so this is a known trade-off rather than a hidden flaw.\n\nWho this is for: a reader working on EEG decoding or cross-modal 3D generation will want to see the architecture and the evaluation pitfalls. It deserves a serious referee, not a desk reject, because the task is important and the approach is coherent. But my own verdict is skeptical until the authors add a proper split, a control condition, and fix the dataset citation.\n\nMy recommendation: engage with it as a conditional submission. Ask for major revision: report held-out accuracy with subject independence, add a no-condition baseline, and correct the data provenance. If those fixes land, this could become a real contribution.","headline":"A plausible first step toward EEG-to-3D, but the evidence is not yet there: no held-out split, no control, and a wrong dataset citation.","tokens_in":11887,"tokens_out":1709,"would_cite":false,"duration_ms":19466,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A dual self-attention EEG encoder can condition a NeRF through variational score distillation to generate 3D objects whose content and structure match the original visual stimuli.","keywords":["EEG decoding","3D object reconstruction","neural radiance fields","variational score distillation","stable diffusion prior","dual self-attention","brain-computer interface","cross-modal generation"],"falsifier":"Compare the same pipeline against a strict held-out split of subjects or object categories, and score the 3D output with a geometric metric such as Chamfer distance between the reconstructed surface and the ground-truth object, or depth-map accuracy against the rendered stimulus. If PSNR falls to chance-level or the geometric metrics are no better than a generic mean-object prior, the claim that EEG drives genuine 3D reconstruction is not supported.","tokens_in":10777,"feed_emoji":"🧠","tokens_out":6807,"duration_ms":69525,"temperature":0.7,"pith_summary":"This paper reports reconstructing 3D visual objects directly from EEG brain signals, a step past the usual EEG-to-2D image reconstruction task. The authors reason that the brain encodes spatial structure even when a person views flat images, so a successful decoder should recover 3D form rather than discard depth. They build a pipeline whose EEG encoder uses a dual self-attention mechanism and is trained with masked self-supervision, contrastive learning, and cross-attention, and whose 3D generator uses Stable Diffusion as a prior to train a NeRF through variational score distillation. On frames rendered from the NeRF and aligned to the original stimulus viewpoints, the pipeline reports better PSNR, LPIPS, and FID than several EEG-to-2D baselines evaluated the same way. If the claim holds, EEG contains enough visual-semantic information to condition volumetric generation, which would matter for brain-computer interfaces and assistive communication.","feed_headline":"EEG signals now reconstruct 3D objects, not just 2D images","feed_subtitle":"An EEG encoder plus diffusion-guided NeRF turns brain recordings into viewable 3D objects.","key_machinery":"The load-bearing mechanism is the EEG embedding $\\xi^e_i$ extracted by the dual self-attention encoder. The first attention layer operates over electrode positions to form EEG tokens; a second layer, augmented by a record token, folds those tokens into a single condition vector. That vector is passed through the Q/K/V projections of cross-attention into a fine-tuned Stable Diffusion U-Net, and variational score distillation (VSD) updates the NeRF parameters $\\theta$ using the difference between the pretrained diffusion score and a LoRA-adapted score for rendered images, as in equation (10). The dual self-attention is what turns noisy multichannel EEG into a stable condition, and VSD is what lets a 2D diffusion prior supervise the 3D radiance field.","core_discovery":"The central claim is that 3D-Telepathy establishes an end-to-end mapping from raw EEG recordings to 3D objects. A dual self-attention encoder first turns multichannel, multi-timepoint EEG into a single condition embedding; that embedding is injected into the cross-attention layers of a fine-tuned Stable Diffusion U-Net, and variational score distillation then optimizes a NeRF so that its rendered views match the distribution of images the diffusion model thinks are plausible for that EEG condition. The paper reports PSNR 28.5203, LPIPS 0.3178, and FID 57.3326 on aligned rendered frames, the best among the EEG-to-2D systems it compares against. The authors take this as evidence that the spatial information in EEG that is lost in 2D-only reconstructions can be recovered in 3D, while acknowledging that sides of an object never visible in the original stimulus are not intelligently filled in.","pith_inferences":["Subject-independent held-out evaluation with true 3D metrics would test whether the reported gains survive generalization; the current 2D-frame protocol cannot by itself distinguish decoding from dataset memorization.","The same encoder-plus-diffusion-prior recipe could be applied to other neural modalities or to video stimuli, where temporal EEG dynamics may encode motion and depth cues not present in still-image training.","Since the ablations show the largest drops when the masked self-supervised and contrastive stage is removed, a simpler contrastive-only encoder might approach the same performance with less training cost.","If EEG truly encodes spatial layout, reconstructing 3D from EEG could serve as a neural probe for how the visual system represents depth and viewpoint; this setup offers a testbed for that question."],"forward_implications":["The mapping from EEG to 3D is end-to-end: one trained encoder feeds a NeRF, so no manual image-to-3D step is needed after training.","Because the condition vector is injected into cross-attention layers of a diffusion U-Net, the same EEG embedding supports 2D rendering supervision during NeRF optimization.","On viewpoint-aligned frames, the pipeline outperforms existing EEG-to-2D baselines on PSNR, LPIPS, and FID, indicating the 3D output retains image-level fidelity.","The visible sides of an object are reconstructed from the EEG; unseen sides are not filled in, so the method recovers observed structure but does not hallucinate.","A brain-computer interface could present a user's observed stimulus as a navigable 3D object rather than a flat image."],"supporting_citations":[{"why":"Supplies the variational score distillation objective and update rule used to train the NeRF.","marker":"[8]"},{"why":"Establishes the approach of training NeRF with a pretrained 2D diffusion prior, extended here to EEG conditioning.","marker":"[9]"},{"why":"Defines the neural radiance field representation that is optimized to produce the 3D object.","marker":"[11]"},{"why":"Provides the Stable Diffusion / latent diffusion prior that scores the rendered images.","marker":"[4]"},{"why":"Supplies the EEG-image pairs used to train the encoder and to align evaluation frames.","marker":"[12]"},{"why":"Baseline EEG-to-2D decoder whose contrastive encoder design is extended with the hybrid training strategy.","marker":"[16]"},{"why":"Baseline EEG-to-2D diffusion method compared in the experiments.","marker":"[2]"},{"why":"Shows contrastive alignment of EEG embeddings with image embeddings, a precursor to the training loss.","marker":"[20]"},{"why":"Provides the LoRA adaptation used to estimate the score of rendered images during VSD.","marker":"[13]"},{"why":"Inspires the masked self-supervised learning stage for the EEG encoder.","marker":"[19]"}],"fun_headline_variants":["EEG to 3D: Brain signals now reconstruct objects","3D objects from EEG: A leap beyond 2D","From EEG to 3D: New method creates viewable objects","EEG decoder produces 3D models, not just pictures"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central claim rests on the assumption that the EEG recordings and target images are genuinely paired and cleanly separated into training and evaluation, so the encoder must decode visual content rather than memorize the stimulus set; if that pairing or split is unreliable, the reported gains could shrink or disappear under a held-out test.","fun_headline_variants_meta":{"raw":{"variants":["EEG to 3D: Brain signals now reconstruct objects","3D objects from EEG: A leap beyond 2D","From EEG to 3D: New method creates viewable objects","EEG decoder produces 3D models, not just pictures"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000353,"raw_usage":{"total_tokens":1941,"prompt_tokens":982,"completion_tokens":959,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":598,"completion_tokens_details":{"reasoning_tokens":887}},"tokens_in":598,"tokens_out":959,"duration_ms":10386,"temperature":1.0,"reasoning_tokens":887,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T22:17:05.429727+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compare the same pipeline against a strict held-out split of subjects or object categories, and score the 3D output with a geometric metric such as Chamfer distance between the reconstructed surface and the ground-truth object, or depth-map accuracy against the rendered stimulus. If PSNR falls to chance-level or the geometric metrics are no better than a generic mean-object prior, the claim that EEG drives genuine 3D reconstruction is not supported.","supporting_citations":[{"cited_title":"Prolific- dreamer: High-fidelity and diverse text-to-3d generation with variational score distillation,","cited_arxiv_id":null,"evidence_quote":"Supplies the variational score distillation objective and update rule used to train the NeRF."},{"cited_title":"Nerf: representing scenes as neural radiance fields for view synthesis,","cited_arxiv_id":null,"evidence_quote":"Defines the neural radiance field representation that is optimized to produce the 3D object."},{"cited_title":"High-resolution image synthesis with latent diffusion models,","cited_arxiv_id":null,"evidence_quote":"Provides the Stable Diffusion / latent diffusion prior that scores the rendered images."},{"cited_title":"MOABB: trustworthy algorithm bench- marking for bcis,","cited_arxiv_id":null,"evidence_quote":"Supplies the EEG-image pairs used to train the encoder and to align evaluation frames."},{"cited_title":"Visual decoding and reconstruction via eeg embeddings with guided diffusion,","cited_arxiv_id":null,"evidence_quote":"Baseline EEG-to-2D decoder whose contrastive encoder design is extended with the hybrid training strategy."},{"cited_title":"Decoding Natural Images from EEG for Object Recognition,","cited_arxiv_id":null,"evidence_quote":"Shows contrastive alignment of EEG embeddings with image embeddings, a precursor to the training loss."},{"cited_title":"Masked au- toencoders are scalable vision learners,","cited_arxiv_id":null,"evidence_quote":"Inspires the masked self-supervised learning stage for the EEG encoder."}],"review_version":1}